xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/m68k/README (revision ca453df649ce9db45b64d73678ba06cbccf9aa11)
1Copyright 2001, 2003, 2004 Free Software Foundation, Inc.
2
3This file is part of the GNU MP Library.
4
5The GNU MP Library is free software; you can redistribute it and/or modify
6it under the terms of the GNU Lesser General Public License as published by
7the Free Software Foundation; either version 3 of the License, or (at your
8option) any later version.
9
10The GNU MP Library is distributed in the hope that it will be useful, but
11WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
12or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
13License for more details.
14
15You should have received a copy of the GNU Lesser General Public License
16along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.
17
18
19
20
21
22                      M68K MPN SUBROUTINES
23
24
25This directory contains mpn functions for various m68k family chips.
26
27
28CODE ORGANIZATION
29
30	m68k             m68000, m68010, m68060
31	m68k/mc68020     m68020, m68030, m68040, and CPU32
32
33
34The m5200 "coldfire", which is m68000 less a few instructions, currently has
35no assembler code support.
36
37
38STATUS
39
40The code herein is old and poorly maintained.  If somebody really cared, it
41could be optimized substantially.  For example,
42
43* mpn_add_n and mpn_sub_n could, with more unrolling be improved from 6 to
44  close to 4 c/l (on m68040).
45
46* The multiplication loops could be sped up by using the FPU.
47
48* mpn_lshift by 31 should use the special-case mpn_rshift by 1 code, and
49  vice versa mpn_rshift by 31 use the special lshift by 1, when operand
50  overlap permits.
51
52* On 68000, mpn_mul_1, mpn_addmul_1 and mpn_submul_1 could check for a
53  16-bit multiplier and use two multiplies per limb, not four.
54
55  Similarly various other _1 operations like mpn_mod_1, mpn_divrem_1,
56  mpn_divexact_1, mpn_modexact_1c_odd.
57
58* On 68000, mpn_lshift and mpn_rshift could use a roll and mask instead of
59  lsrl and lsll.  This promises to be a speedup, effectively trading a 6+2*n
60  shift for one or two 4 cycle masks.  Suggested by Jean-Charles Meyrignac.
61
62* config.guess detects 68000, 68010, CPU32 and 68020 by running some code,
63  but relies on system information for 030, 040 and 060.  Can they be
64  identified by running some code?  Currently this only makes a difference
65  to the compiler options selected, since we have no specific asm code for
66  those chips.
67
68One novel idea for 68000 would be to use a 16-bit limb instead of 32-bits.
69This would suit the native 16x16 multiply, but might make it difficult to
70get full value from the native 32x32 add/sub/etc.  This would be an ABI
71option, and would select "__GMP_SHORT_LIMB" in gmp.h.
72
73Naturally an entirely new set of asm subroutines would be needed for a
7416-bit limb.  Also there's various places in the C code assuming limb>=long,
75which would need to be updated, eg. mpz_set_ui.  Some of the nails changes
76may have helped cover some of this.
77
78
79ASM FILES
80
81The .asm files are put through m4 for macro processing, and with the help of
82configure give either MIT or Motorola syntax.  The generic mpn/asm-defs.m4
83is used, together with mpn/m68k/m68k-defs.m4.  See comments in those files.
84
85Not all possible syntax variations are covered.  GCC config/m68k for
86instance has things like $ for immediates on CRDS or reversed cmp order for
87AT&T SGS.  These could probably be handled if anyone really needs it.
88
89
90CALLING CONVENTIONS
91
92The SVR4 standard has an int of 32 bits, and all parameters 32-bit aligned
93on the stack.
94
95PalmOS and perhaps various embedded systems intended for 68000 however use
96an int of 16 bits and parameters only 16-bit aligned on the stack.  This is
97generated by "gcc -mshort" (and is the default for the PalmOS gcc port, we
98believe).
99
100The asm files adapt to these two ABIs by checking sizeof(unsigned), coming
101through config.m4 as SIZEOF_UNSIGNED.  Only mpn_lshift and mpn_rshift are
102affected, all other routines take longs and pointers, which are 32-bits in
103both cases.
104
105Strictly speaking the size of an int doesn't determine the stack padding
106convention.  But if int is 16 bits then we can definitely say the host
107system is not SVR4, and therefore may as well assume we're in 16-bit stack
108alignment.
109
110
111REFERENCES
112
113"Motorola M68000 Family Programmer's Reference Manual", available online,
114
115	http://e-www.motorola.com/brdata/PDFDB/docs/M68000PM.pdf
116
117"System V Application Binary Interface: Motorola 68000 Processor Family
118Supplement", AT&T, 1990, ISBN 0-13-877553-6.  Has details of calling
119conventions and ELF style PIC coding.
120
121
122
123----------------
124Local variables:
125mode: text
126fill-column: 76
127End:
128